Rebuild, Regenerate: A Gated Temporal Convolution Based Gan for Speech Signal Improvement

This paper describes the Sluice Team’s submission to the ICASSP 2024 Signal Improvement Challenge. We have developed an efficient single stage network to improve speech signal quality degraded due to several distortions introduced in real-time communication. Our concept is to Rebuild and Regenerate the signal using GAN(Generative Adversarial Network). In our proposed system, the generator is used to rebuild information from distorted signals and the adversarial training is used to regenerate the completely lost information. The official results show that the proposed method ranks 4th in the Challenge with a overall MOS score of 2.895 and word accuracy of 0.8037.